
A cat on a crowded shelf has one method for making room. A pen arrives, then a mug, then your keys, and each time something older goes over the edge. Usually the oldest thing. Usually without the cat looking.
An AI model’s context window works much the same way, minus the eye contact. It is the amount of text the model can take into account at once when it writes its next reply. Everything has to fit on that shelf together: the instructions the app gives the model behind the scenes, the whole conversation so far, any documents you pasted in, and the reply it is in the middle of writing.
How big is the shelf
The size is counted in tokens rather than words. A token is a chunk of text, often a short word, sometimes part of a longer one. For English, a fair rule of thumb is that 100 tokens come to about 75 words. Current tools range from windows of a few thousand tokens to several hundred thousand, which in the bigger cases is more than a long novel.
That sounds like plenty, and for a quick question it is. But a long working session fills it faster than you would think. Paste in a few files, go back and forth for an hour, and the shelf is crowded.
What happens when it is full
That depends on the app. Some quietly drop the oldest messages. Some summarise the early part of the conversation into a short note and carry on from that, which is often called compaction. Some stop and ask you to start a new chat.
In none of these cases is the model forgetting the way a person forgets. It has no memory of its own between replies; each time, the app sends the conversation again from the top. If a message is no longer in what gets sent, it is simply not there. (Some apps add a separate memory feature that saves notes between chats. That is a different mechanism with its own quirks.)
There is a subtler problem as well. Even inside the window, very long inputs are not used evenly. A detail buried in the middle of a long document is more likely to be missed than one near the start or the end.
Why people who use it find it funny
Anyone who has spent an afternoon on a hard problem with an AI assistant knows the moment. The careful decisions from the first hour have been squashed into a line. The assistant brightly suggests the very approach you both ruled out at lunchtime. It is not being stubborn. It can no longer see lunchtime.
That is the cat on the shelf. Calm, efficient, faintly pleased with itself, and entirely unbothered by what just hit the floor.
The shirts
Context Compacted shows exactly that step: a long scroll of tiny writing, an arrow, and a yellow sticky note with the whole thing boiled down. It is a €26 tee and also a hoodie at €49 to €53. For December there is All I Want for Christmas Is a Larger Context Window, a small robot in a Santa hat holding out a present. It had a longer wish list. It cannot remember the rest.
Related: why AI agents hand work to subagents with their own fresh context, a gift guide for people who work with AI all day, and the Christmas collection.
